Labor Automation

Between 1920 and 1940, AT&T converted American cities from manual switchboards to dial service. In converted cities, young women's employment as telephone operators fell by 50 to 80 percent.

Women already working as operators were more likely to end up in lower-paying jobs or out of work. Later cohorts found work as typists, secretaries, and in service jobs, and overall employment of young women did not fall.^[1]

The work of connecting callers continued, performed by mechanical switches. The operators' labor, the hours they sold, was displaced.

Bank tellers went through a slower version of the same change. Automated teller machines were introduced in the 1970s, and their number in the United States quadrupled from about 100,000 to 400,000 between 1995 and 2010. Teller employment rose anyway, from about 500,000 in 1980 to about 550,000 in 2010, though more slowly than total employment.^[2]

ATMs took over the cash-handling steps of the job. Each branch needed fewer tellers, but ATMs made branches cheaper to run, and with help from deregulation the number of urban branches rose by more than 40 percent between 1988 and 2004. Banks came to use tellers as salespeople who introduced customers to credit cards, loans, and investment products.^[2]

Later technology took over more of the job. Customers now deposit checks with their phones, and banks have closed branches. The Bureau of Labor Statistics counted about 339,200 teller jobs in 2025 and projects a 13 percent decline by 2035.^[3]

The operators lost nearly every step of their job to dial service within a few years. Machines have taken over the teller's job a few steps at a time over fifty years, and for decades the tellers who stayed spent more of their time on the steps the machines could not do.

Defining Automation

Defining automation can be tricky, because it is applied in different ways and different professions define it differently.

Engineers describe an automated system as one in which "a process is performed by a machine without the direct participation of a human worker."^[4] The word itself comes from Ford. In April 1947, Delmar Harder, who headed Ford's manufacturing division, set up an Automation Department to automate the transfer of materials from one operation to the next.^[5]

Human-factors researchers treat automation as a matter of degree. Raja Parasuraman, Thomas Sheridan, and Christopher Wickens define it as the full or partial replacement of a function that a person used to carry out. They divide any task into four functions:

1. acquiring information
2. analyzing it
3. choosing a decision or action
4. carrying the action out

Each function can be automated to a different degree. For decisions and actions, they use a ten-level scale that runs from a computer that "offers no assistance" to one that "decides everything, acts autonomously, ignoring the human."^[6]

• Retain (level 1): the computer offers no assistance.
• Augment (levels 2 to 5): the computer offers all the options, narrows them to a few, suggests one, or acts on its suggestion if the person approves.
• Automate (levels 6 to 10): the computer acts unless the person vetoes in time, acts and then informs the person, informs the person only if asked or only if it decides to, or decides everything.

At levels 2 through 5, the computer helps and the person decides. In the authors' example, a system that warns an air-traffic controller of a conflict between two flight paths and suggests a resolution is at level 4. At level 6 or higher, it would carry out its own resolution unless the controller stepped in.^[6]

Economists define automation as capital performing a task that labor used to perform.^[7]

Augmentation is the other side of the same choice. David Autor and his coauthors classify innovations by what they touch. Automation innovations substitute for what workers do and reduce labor demand in the occupations they reach. Augmentation innovations raise the value of what an occupation produces, and they create new work.^[8]

For a single step in a task, the choice is to retain it, augment it with a tool, or automate it. On the scale above, retaining a step is level 1, augmenting it is levels 2 through 5, and automating it is level 6 and higher.

Augmentation has been a goal of computing since its early days. In 1960, J. C. R. Licklider described a "man-computer symbiosis" in which people "will set the goals, formulate the hypotheses, determine the criteria, and perform the evaluations," and computers do "the routinizable work."^[9]

Two years later, Douglas Engelbart defined "augmenting human intellect" as "increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems."^[10]

Seen from the job, most automation is partial: machines take over some steps, and a person keeps the rest. James Bessen traced the 271 detailed occupations in the 1950 census to 2010 and found only one, elevator operator, whose disappearance could be largely attributed to automation. He also found that occupations that used computers more grew faster, even among routine and middle-wage occupations, though much of that growth came at the expense of other occupations.^[11]

Autor has argued that commentators "tend to overstate the extent of machine substitution for human labor and ignore the strong complementarities between automation and labor that increase productivity, raise earnings, and augment demand for labor."^[2] The two kinds of innovation also tend to reach the same jobs: occupations heavily exposed to automating patents are usually heavily exposed to augmenting patents as well.^[8] Sebastian Raisch and Sebastian Krakowski reach a similar conclusion about AI in management, where "augmentation cannot be neatly separated from automation."^[12]

Both kinds of change run through physical and cognitive labor:

• After 1826, as power looms spread in Britain, the piece rates paid to handloom weavers collapsed, and by the 1840s hand-weaving of cotton had largely disappeared.^[13]
• Farm machinery augmented the farmers who stayed on the land. Growing 100 bushels of wheat took 250 to 300 hours of labor in 1830 and 3 to 4 hours in 1975, with large tractors and combines that a farmer still drove, as Part 1's "Structural Transformation" described.^[14]
• Computers took over calculation and record-keeping that followed explicit rules, and they complemented the people who solved problems and handled complex communication.^[15]
• LLMs now reach steps whose output is open text.

Automation Acts on Steps

Automating a step moves some or all of its hours to a machine. With augmentation, the person keeps the step and it takes less time. For a step that took a given number of hours before the change:

Labor productivity for the task is multiplied by the ratio of total hours before to total hours after.

Automating a step often adds a new one, because someone has to check what the machine produced.

Take the claims workflow from "The Task Structure of Work." Suppose a firm automates intake, the policy lookup, and evidence gathering, so that a machine does 90 percent of that work and the analyst handles the exceptions. The firm gives the analyst a tool that makes payout estimates and fraud screening 25 percent faster, and it leaves the coverage check and the final decision as they were. Someone also has to check the extracted evidence against the source documents.

Automate Augment Retain New human step added by automation

Step	Decision	Before	After
1. Intake and classify	Automate	10 h	1 h
2. Pull policy and terms	Automate	10 h	1 h
3. Gather evidence	Automate	40 h	4 h
+ Check evidence	New step	0 h	12 h
4. Check coverage	Retain	10 h	10 h
5. Estimate payout	Augment	15 h	12 h
6. Screen for fraud	Augment	10 h	8 h
7. Decide and communicate	Retain	5 h	5 h
Total		100 h	53 h

For every 100 hours the task took before, it now takes 53, so the analyst can handle about 1.9 times as many claims per hour. Without the checking step the task would take 41 hours, and the estimate would be 2.4 times. Checking takes back about a fifth of the hours that automation saved.

The mix of the job changes as well. Before the change, 40 percent of the analyst's hours went to the coverage check, the payout estimate, fraud screening, and the decision. Afterward about two-thirds do, and most of the rest goes to checking. In "The Value of Human Labor," those four steps carried 80 percent of the value of the claim.

The hours freed by automation go to one of three places:

• other steps in the same task, which the worker can now do more carefully
• other tasks in the same role, such as complex or disputed claims
• out of the role, as fewer hours bought or fewer people hired, which is displacement

Which of these dominates decides whether the role grows, changes, or shrinks.

Quality matters as much as hours. In a chain of dependent steps, a weak automated step lowers the quality of the whole output, so a credible estimate discounts the hours saved by any loss of quality.^[16] Joshua Gans and Avi Goldfarb show that automating one step changes the return to automating the others, so adoption tends to arrive in bundles.^[17]

For a CTO, the business case for automating a step needs four numbers:

• the hours removed
• the hours of checking added
• the change in the quality of the output
• where the freed hours will go

The Limits of Automation

Automation stops in a few predictable places:

• Steps whose rules cannot be written down. People know more than they can tell, and such steps resisted automation until machine learning could learn some of them from examples.^[18] LLMs learned many text steps this way, which moved this limit for text.
• Output that is costly to check. A payroll system can be verified once and trusted for every run. An analysis written by an LLM has to be checked case by case, and checking it takes the same kind of judgment as writing it. Parasuraman and his coauthors advise against high levels of decision automation where errors are costly and a person may have to take over when the automation fails.^[6]
• Decisions someone has to answer for. A person can be held responsible for a decision, and a model cannot.^[19] European law writes part of this down. The GDPR gives people the right not to be subject to a decision "based solely on automated processing" that significantly affects them. The EU AI Act requires that high-risk AI systems "can be effectively overseen by natural persons," who must be able to override them.^[20]
• Work that costs more to automate than to staff. Only about 23 percent of the wages paid for tasks exposed to computer vision would be cost-effective to automate at recent costs.^[21]

For LLMs, the first limit binds less than it did for earlier software, and the second and third bind more. Both keep a person in the task even where an LLM can perform most of its steps. They are also why an LLM agent can perform the steps of a role while a person still holds the job and answers for the result, as "What Is Work?" described.

On whole projects, agents still fall short most of the time. On the Remote Labor Index, a set of 240 paid freelance projects, evaluators judged the best AI agent's work acceptable on 2.5 percent of the projects in October 2025 and on 15.8 percent by July 2026.^[22] The rate is rising quickly.

Oversight is labor in its own right. Lisanne Bainbridge described the problem in 1983, writing about the operators of automated industrial processes:

Bainbridge also pointed out that people cannot keep watching a system where little happens for more than about half an hour. Skills also fade when they are not used, so "a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one."^[23]

Raja Parasuraman and Dietrich Manzey later reviewed two failure modes in people who supervise automation. One is complacency: people catch fewer of a system's errors than they would if they did the work themselves. The other, automation bias, is taking the machine's output as a shortcut for looking up and weighing the evidence. Both show up in experts as well as novices, and training does not prevent them.^[24] The EU AI Act names automation bias as something the people overseeing high-risk systems must stay aware of.^[20]

The same costs appear with LLMs. In a randomized trial in early 2025, 16 experienced open-source developers took 19 percent longer to finish tasks in their own projects when they were allowed to use AI tools. They had expected to be 24 percent faster, and afterward they believed they had been about 20 percent faster. They spent about 9 percent of their time reviewing and cleaning up the AI's output. The tools have improved since then, and a 2026 follow-up by the same group found signs of speedups, which the group itself considers unreliable.^[25]

An honest estimate of automation's savings counts the hours of checking it adds. Part 9 examines that verification work.

LLMs Lean Toward Augmentation

An LLM can draft a summary of a claim, pull the facts out of a file, or propose a payout. In most knowledge work today, a person still reads the result, corrects it, and signs off on it. On the scale earlier in this article, that is level 4 or 5: the machine suggests and the person decides. The drafting step is automated, and the task as a whole is augmented.

Usage data show the same pattern. In conversations from December 2024 and January 2025, Anthropic classified 57 percent of the use of its Claude models as augmentation, where the person refined the work with the model, used it to learn, or asked it to check their work, and 43 percent as automation, where the model carried out the task with little back-and-forth.^[26]

Later reports, measured on a slightly different basis, show the split moving in both directions. Automation briefly overtook augmentation in August 2025, and by February 2026 augmentation was back to 53 percent against 44 percent for automation. On Anthropic's API, where businesses build the models into their own software, most use is automation, though that share fell from 77 percent to 68 percent over the same six months.^[27] An API call often comes from a program that hands the model a single step, so the exchange looks like automation even when a person reviews the result afterward, which these data do not record.

In Anthropic's study of software development, 79 percent of conversations with Claude Code, an agent that works inside a developer's environment, were automation, compared with 49 percent of coding conversations in the Claude chat app.^[28] Code can be run against tests, which makes it cheaper to check than most text.

The experiments that found large gains from LLMs tested augmentation, where a person used the tool and kept responsibility for the output. Professional writers finished tasks about 40 percent faster, and customer-support agents resolved about 15 percent more issues per hour, with the largest gains for the least experienced workers, as "The Value of Human Labor" described.^[29][30] Among consultants using GPT-4, Fabrizio Dell'Acqua and colleagues found two patterns of successful use. "Centaurs" divided the work between themselves and the AI, and "cyborgs" integrated the AI into their own work and interacted with it continually.^[31]

In labor-market data the effects have been small so far. Anders Humlum and Emilie Vestergaard linked surveys of Danish workers to administrative records and found no measurable effect of chatbot adoption on earnings or recorded hours. Adopters reported saving about 3 percent of their work hours. About 8 percent of chatbot users had taken on entirely new tasks, and about a third of the new AI-related work went to reviewing AI output for quality and compliance.^[32]

The clearest early signal is at entry level. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen estimate that employment of 22- to 25-year-olds in the most AI-exposed occupations is about 19 percent below where it would be had it kept pace with less-exposed peers, mostly because firms hired fewer of them. The declines are concentrated in occupations where AI is used mostly to substitute for human tasks. Where it is used mostly to complement workers, employment is flat or rising. The authors call these early, descriptive indicators, and the patterns weaken when they control for education.^[33]

These results fit the step model. Steps are being automated inside jobs that persist. Part of the time saved goes to checking, and hiring adjusts before headcount does.

Firms can still choose automation where augmentation would produce more. Daron Acemoglu, Autor, and Simon Johnson argue that firms have reasons to automate beyond productivity, since automation weakens workers' bargaining position, and that AI developers have built their roadmaps around replacing human expertise.^[34] Erik Brynjolfsson calls the focus on building AI that imitates and replaces people "the Turing Trap." He argues that augmentation "creates new capabilities and new products and services, ultimately generating far more value than merely human-like AI."^[35]

The gap between large gains in experiments and small effects in labor markets comes partly from measuring at different levels.

Where the Savings Go

Automation and augmentation both cut the labor needed for each unit of output. In the claims example, the same claims take about half the analyst hours. What happens next depends on what the firm does with the savings and on how its customers respond.

In 1821, David Ricardo added a chapter on machinery to the third edition of his Principles. He conceded that workers' fear of machinery "is not founded on prejudice and error, but is conformable to the correct principles of political economy."^[42] Keynes described the risk in 1930 as "unemployment due to our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour."^[43]

Daron Acemoglu and Pascual Restrepo describe three ways technology changes the demand for labor:^[44]

• the displacement effect: capital takes over tasks that labor used to perform
• the productivity effect: lower costs raise output, and with it the demand for labor in the tasks that are not automated
• the reinstatement effect: new tasks bring labor back into a broader range of work

They argue that the technologies that threaten employment and wages are "so-so technologies" that generate small productivity improvements. These displace workers while producing little of the productivity effect.^[44]

Augmentation does not settle the question either. Farm machinery made each farmer far more productive, yet agriculture's share of the US workforce fell from 41 percent in 1900 to 2 percent in 2000, as the share of household income spent on food declined.^[2] Autor and his coauthors find that over the past four decades, the losses in labor demand from automation innovations have grown, while the gains from augmentation innovations have not.^[8]

The two cases from the opening ended differently. AT&T raised its rates after automating switching.^[1] ATMs made branches cheaper to run, and banks opened more of them.^[2]

Whether the hours freed by LLMs come back as new work depends on whether the savings reach buyers as lower prices, how much more buyers want when prices fall, and whether new goods and services appear. "Demand," the next article in Part 5, takes up those questions.

---

References

1. Feigenbaum, J., & Gross, D. P. (2024). Answering the call of automation: How the labor market adjusted to mechanizing telephone operation. Quarterly Journal of Economics, 139(3), 1879–1939. doi:10.1093/qje/qjae005 [rate increases after dial cutovers: Appendix Figure A.4]
2. Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30. doi:10.1257/jep.29.3.3 [complementarities and agriculture's share of employment, p. 5; ATMs and bank tellers, citing Bessen (2015), pp. 6–7; food's share of household income, p. 7]
3. U.S. Bureau of Labor Statistics. Tellers. Occupational Outlook Handbook. Last modified August 27, 2026. bls.gov/ooh/office-and-administrative-support/tellers.htm [about 339,200 jobs in 2025; projected change 2025–35, −13 percent]
4. Groover, M. P. (2016). Automation, Production Systems, and Computer-Integrated Manufacturing (4th ed., Global Edition). Pearson, p. 20.
5. Hounshell, D. A. (1995). Ford automates: Technology and organization in theory and practice. Business and Economic History, 24(1), 59–71, at p. 69. thebhc.org [John Diebold's Automation: The Advent of the Automatic Factory (Van Nostrand, 1952) popularized the word.]
6. Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30(3), 286–297. doi:10.1109/3468.844354 [definition, Table I, and the air-traffic example, p. 287; high-cost decisions, p. 292] Table I builds on Sheridan, T. B., & Verplank, W. L. (1978). Human and Computer Control of Undersea Teleoperators. MIT Man-Machine Systems Laboratory.
7. Acemoglu, D., & Restrepo, P. (2018). The race between man and machine: Implications of technology for growth, factor shares, and employment. American Economic Review, 108(6), 1488–1542. doi:10.1257/aer.20160696
8. Autor, D., Chin, C., Salomons, A., & Seegmiller, B. (2024). New frontiers: The origins and content of new work, 1940–2018. Quarterly Journal of Economics, 139(3), 1399–1465. doi:10.1093/qje/qjae008 [Abstract: "The demand-eroding effects of automation innovations have intensified in the past four decades while the demand-increasing effects of augmentation innovations have not." Exposure to automating and augmenting patents is strongly correlated across occupations.]
9. Licklider, J. C. R. (1960). Man-computer symbiosis. IRE Transactions on Human Factors in Electronics, HFE-1(1), 4–11, at p. 4. doi:10.1109/THFE2.1960.4503259
10. Engelbart, D. C. (1962). Augmenting Human Intellect: A Conceptual Framework. Summary Report AFOSR-3223. Stanford Research Institute, p. 1. dougengelbart.org/pubs/augment-3906.html
11. Bessen, J. (2016). How Computer Automation Affects Occupations: Technology, Jobs, and Skills. Boston University School of Law, Law and Economics Working Paper No. 15-49 (revised May 2016). ssrn.com/abstract=2690435 [elevator operators, p. 5; computer use and employment growth, p. 2. The net effect after substitution between occupations is small, about 0.45 percent a year.] See also Bessen, J. (2016, September 22). How computer automation affects occupations: Technology, jobs, and skills. VoxEU/CEPR.
12. Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation–augmentation paradox. Academy of Management Review, 46(1), 192–210, at p. 192. doi:10.5465/amr.2018.0072
13. Allen, R. C. (2018). The hand-loom weaver and the power loom: A Schumpeterian perspective. European Review of Economic History, 22(4), 381–402. doi:10.1093/ereh/hex030 See also Bythell, D. (1969). The Handloom Weavers. Cambridge University Press. doi:10.1017/CBO9780511895999
14. Greenwood, J., & Seshadri, A. (2005). Technological progress and economic transformation. In P. Aghion & S. N. Durlauf (Eds.), Handbook of Economic Growth (Vol. 1B, pp. 1225–1273). Elsevier. doi:10.1016/S1574-0684(05)01019-1
15. Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. Quarterly Journal of Economics, 118(4), 1279–1333. doi:10.1162/003355303322552801
16. Kremer, M. (1993). The O-ring theory of economic development. Quarterly Journal of Economics, 108(3), 551–575. doi:10.2307/2118400
17. Gans, J. S., & Goldfarb, A. (2026). O-Ring Automation. NBER Working Paper 34639. doi:10.3386/w34639
18. Polanyi, M. (1966). The Tacit Dimension. Doubleday. See also Autor, D. H. (2014). Polanyi's Paradox and the Shape of Employment Growth. NBER Working Paper 20485. doi:10.3386/w20485
19. Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., et al. (2018). AI4People—An ethical framework for a good AI society: Opportunities, risks, principles, and recommendations. Minds and Machines, 28(4), 689–707. doi:10.1007/s11023-018-9482-5
20. Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22(1). data.europa.eu/eli/reg/2016/679/oj; Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14(1) and 14(4). data.europa.eu/eli/reg/2024/1689/oj [Article 14(4)(b) names "automation bias"; 14(4)(d) covers the power to "disregard, override or reverse" a system's output. Regulation (EU) 2026/1744 leaves Article 14's wording unchanged and moves its application to 2 December 2027 or 2 August 2028, depending on the type of system.]
21. Svanberg, M., Li, W., Fleming, M., Goehring, B., & Thompson, N. (2024). Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? SSRN Working Paper 4700751. doi:10.2139/ssrn.4700751
22. Mazeika, M., Gatti, A., Menghini, C., et al. (2025). Remote Labor Index: Measuring AI Automation of Remote Work. Scale AI and Center for AI Safety. arXiv:2510.26787 [2.5 percent at launch, October 2025]. Update: Center for AI Safety. (2026, July 1). Significant increase in digital labor automation. safe.ai/blog/significant-increase-in-digital-labor-automation [15.8 percent, the highest rate measured as of that post]
23. Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. doi:10.1016/0005-1098(83)90046-8 [the designer's irony and skill loss, p. 775; attention lasting about half an hour, p. 776]
24. Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. doi:10.1177/0018720810376055 [Abstract: complacency "is found in both naive and expert participants and cannot be overcome with simple practice"; automation bias "cannot be prevented by training or instructions."]
25. Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR. arXiv:2507.09089 [246 tasks, February–June 2025; review and cleanup time from the time-use analysis]. Follow-up: METR. (2026, February 24). We are changing our developer productivity experiment design. metr.org/blog/2026-02-24-uplift-update
26. Handa, K., Tamkin, A., et al. (2025). Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations. arXiv:2503.04761. [figures from the 2025 arXiv version; Claude.ai Free and Pro conversations from December 2024 and January 2025]
27. Anthropic. (2025, September 15). Anthropic Economic Index report: Uneven geographic and enterprise AI adoption. anthropic.com/research/anthropic-economic-index-september-2025-report; and Anthropic. (2026, March 24). Anthropic Economic Index report: Learning curves. anthropic.com/research/economic-index-march-2026-report [Claude.ai: 47.0 percent augmentation and 49.1 percent automation in August 2025, 53 and 44 percent in February 2026; first-party API automation 77 percent in August 2025 and 68 percent in February 2026. These are shares of all conversations, so they do not sum to 100 percent and are not directly comparable with the 57/43 split.]
28. Anthropic. (2025). Anthropic Economic Index: AI's Impact on Software Development. anthropic.com/research/impact-software-development
29. Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. doi:10.1126/science.adh2586
30. Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889–942. doi:10.1093/qje/qjae044
31. Dell'Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013. doi:10.2139/ssrn.4573321 Published in revised form in Organization Science, 37(2), 403–423 (2026), doi:10.1287/orsc.2025.21838. The "centaur" and "cyborg" descriptions are from the working paper.
32. Humlum, A., & Vestergaard, E. (2025; revised March 2026). Large Language Models, Small Labor Market Effects. NBER Working Paper 33777. doi:10.3386/w33777 The March 2026 revision is retitled Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. [figures from the March 2026 revision: time savings of about 3 percent of work hours; about 8 percent of users with entirely new tasks; 35 percent of new AI-related work in AI quality review and AI ethics and compliance; effects larger than 2 percent ruled out]
33. Brynjolfsson, E., Chandar, B., & Chen, R. (2025; revised August 2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab working paper. digitaleconomy.stanford.edu [The 19% figure is the August 2026 version's descriptive "kept-pace" shortfall, using data through June 2026. Substitution versus complementary use is the August 2026 version's Fact 5, measured with Anthropic Economic Index data.]
34. Acemoglu, D., Autor, D., & Johnson, S. (2026). Building Pro-Worker Artificial Intelligence. The Hamilton Project, Brookings Institution. brookings.edu
35. Brynjolfsson, E. (2022). The Turing Trap: The promise & peril of human-like artificial intelligence. Daedalus, 151(2), 272–287, at p. 272. doi:10.1162/daed_a_01915
36. Solow, R. M. (1987, July 12). We'd better watch out. New York Times Book Review, p. 36.
37. Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333–372. doi:10.1257/mac.20180386
38. Solow, R. M. (1957). Technical change and the aggregate production function. Review of Economics and Statistics, 39(3), 312–320. doi:10.2307/1926047
39. Autor, D., & Salomons, A. (2018). Is automation labor share–displacing? Productivity growth, employment, and the labor share. Brookings Papers on Economic Activity, Spring 2018, 1–87. doi:10.1353/eca.2018.0000
40. Acemoglu, D., & Restrepo, P. (2020). Robots and jobs: Evidence from US labor markets. Journal of Political Economy, 128(6), 2188–2244. doi:10.1086/705716
41. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308. doi:10.1126/science.adj0998
42. Ricardo, D. (1821). On the Principles of Political Economy and Taxation (3rd ed.). John Murray. Chapter 31, "On Machinery."
43. Keynes, J. M. (1930). Economic possibilities for our grandchildren. The Nation and Athenæum, 11 and 18 October 1930. Reprinted in Essays in Persuasion (1931). Macmillan.
44. Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30. doi:10.1257/jep.33.2.3 [the three effects, pp. 3–4; "so-so technologies," pp. 10 and 26]